A Comparative Study of Deep Learning Semantic Segmentation Models for Kidney Segmentation in Ultrasound Images Using the Open Kidney Ultrasound Dataset

Accurate segmentation of kidney ultrasound (KUS) images is essential for various diagnostic procedures. However, challenges such as variability in image quality and the need for real-time processing remain significant obstacles. This study presents a comparative evaluation of eleven well-established deep learning (DL) semantic segmentation architectures, including SegFormer, U-Net, HED-UNet, DSEU-Net, UNet++, Multi-scale Attention Net (MA-Net), BiSeNet V2, Pyramid Attention Network (PAN), LinkNet, SegNet, and DeepLabV3+. Different backbones are examined for three architectures (SegFormer, SegNet, and DeepLabV3+), resulting in a total of twenty distinct DL models. The models are trained and evaluated using the publicly available Open Kidney Ultrasound dataset. Additionally, the models are assessed based on three main indicators: kidney segmentation performance measured by six metrics, inference time per KUS image, and statistical differences quantified using the F1-score and intersection over union metrics. To the best of our knowledge, the SegFormer, HED-UNet, MA-Net, BiSeNet V2, and PAN models have not been previously applied to kidney segmentation in KUS images. The results indicate that the SegFormer models with the B5 and B4 MiT backbones have achieved the highest segmentation performance while maintaining interactive processing speeds. By using a publicly available dataset, this study supports reproducibility and standardized benchmarking, helping to overcome challenges associated with the use of private KUS image datasets in earlier studies. The findings highlight the strengths and limitations of each model, providing insights to guide future research and improve clinical applicability of DL-based kidney segmentation in KUS imaging.